States evolution in Θ(λ)-learning based on logical MDPs with negation

Based on the Logical MDPs with Negation, a model of Relational Reinforcement Learning, proposed in [1], we define the self-loop degree and the common characteristic of abstract states formally, and propose an evolution process collaborated with Theta (lambda)-learming according to the formal definitions. The abstract state space will be self-organized in the evolution process rather than given manually by human. The experiments show that the agent can catch the essence of the given task, and the self-organized states are rational.